Semester of Graduation

Summer 2023

Degree

Master of Electrical Engineering (MEE)

Department

The Department of Electrical and Computer Engineering

Document Type

Thesis

Abstract

In a world of continuously advancing technology, the reliance on these technologies continues to increase. Recently, transformer networks [22] have been implemented through various projects such as ChatGPT. These networks are extremely computationally demanding and require cutting-edge hardware to explore. However, with the growing increase and popularity of these neural networks, a question of reliability and resilience comes about, especially as the dependency and research on these networks grow. Given the computational demand of transformer networks, we investigate the resilience of the weights and biases of the predecessor of these networks, i.e. the Long Short-Term (LSTM) neural network, through four implementations of the original LSTM network. Based on the observations made through fault injection of these networks, we propose an effective means of fault mitigation through Hamming encoding of selected weights and biases in a given network and lay the groundwork for similar mitigation methods with transformers.

Date

5-6-2023

Committee Chair

Vaidyanathan, Ramachandran

DOI

10.31390/gradschool_theses.5780

Share

COinS